Towards predicting protein-protein interactions in novel organisms

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning

Background:

• Machine learning is widely used for predicting protein-protein interactions (PPI).
• Current methods often train and evaluate predictive models using data from well-characterized reference organisms.
• This approach is common for inferring potential interactions in novel organisms lacking experimental PPI data.

Purpose of the Study:

• To investigate the validity of evaluating machine learning models for PPI prediction on the same organism used for training.
• To assess the performance of PPI prediction models across genetically distinct organisms.
• To highlight the limitations of current evaluation practices in bioinformatics.

Main Methods:

• Development and evaluation of machine learning models for predicting protein-protein interactions.
• Comparative analysis of model performance using training and testing datasets from both the same and genetically distinct organisms.
• Utilizing known PPI data from reference organisms for model inference and validation.

Main Results:

• Evidence suggests that evaluating models on the same organism used for training provides a misleading indication of real-world performance.
• Model performance significantly differs when applied to genetically distinct organisms compared to the training organism.
• The practice of evaluating models on reference organisms is not a reliable proxy for performance on novel, genetically distinct species.

Conclusions:

• Evaluating machine learning models for protein-protein interaction prediction on the same organism used for training is not scientifically sound.
• This methodology is unsuitable for predicting interactions in novel organisms lacking PPI data.
• There is a critical need to develop and implement evaluation strategies that use organisms genetically distinct from the training set to ensure reliable predictions.